🤖 AI Document Comparison Agent (Agentic LangGraph Version)

Agentic RAG — the AI autonomously decides what to search, how many times, and reasons step-by-step before producing a cited, structured answer.

---

Why "agentic" vs simple RAG?

Simple RAG (v1)	Agentic RAG (v2)
Always runs: retrieve A → retrieve B → LLM	Decides what to retrieve based on the question
Fixed 2-retrieval pipeline	Can call tools 5-10+ times, refining queries each time
One prompt → one answer	Reasoning loop: Thought → Action → Observation → repeat
No awareness of what it doesn't know	Can say "I need to search X with a better query"

---

Architecture

                     ┌─────────────────────────────────┐
                     │     LangGraph ReAct Agent        │
                     │                                  │
User Query ─────────►│  Thought: "Let me first get      │
                     │  an overview of both docs..."    │
                     │            │                     │
                     │     ┌──────▼──────┐              │
                     │     │  Tool Call  │              │
                     │     │ get_doc_    │              │
                     │     │ overview()  │              │
                     │     └──────┬──────┘              │
                     │            │ Observation          │
                     │  Thought: "Now I'll compare      │
                     │  the methodology sections..."    │
                     │            │                     │
                     │     ┌──────▼──────┐              │
                     │     │  Tool Call  │              │
                     │     │ compare_    │              │
                     │     │ topic()     │              │   ◄── Mistral Embed
                     │     └──────┬──────┘              │       ChromaDB A
                     │            │ Observation          │       ChromaDB B
                     │  Thought: "There's a conflict,   │
                     │  let me investigate..."          │
                     │            │                     │
                     │     ┌──────▼──────┐              │
                     │     │  Tool Call  │              │
                     │     │ find_       │              │
                     │     │ conflicts() │              │
                     │     └──────┬──────┘              │
                     │            │                     │
                     │     Final Answer (Mistral Large) │
                     └─────────────────────────────────┘

---

6 Agent Tools

Tool	Purpose	When agent uses it
search_document_a	Semantic search in Doc A	Targeted evidence gathering
search_document_b	Semantic search in Doc B	Targeted evidence gathering
compare_topic	Parallel search in both docs	Primary comparison tool
find_conflicts	Multi-angle contradiction search	Disagreement detection
find_common_ground	Agreement/consensus search	Similarity detection
get_document_overview	Broad topic sweep	Understanding doc scope

---

Quick Start

pip install -r requirements.txt
python -m pip install pdfplumber
python -m pip install langchain-mistralai langchain-chroma
# If you encounter chromadb / OpenTelemetry import errors, pin these versions:
# python -m pip install opentelemetry-api==1.41.1 opentelemetry-sdk==1.41.1 opentelemetry-exporter-otlp-proto-grpc==1.41.1

streamlit run app.py

Then in the UI:

1. Enter your Mistral API key (sidebar)
2. Upload PDF A and PDF B
3. Ask a question — watch the agent reason live

---

File Structure

doc_compare_agent/
├── app.py             # Streamlit UI with live agent step rendering
├── agent.py           # LangGraph ReAct agent + streaming
├── agent_tools.py     # 6 LangChain tools
├── vector_store.py    # LangChain-Chroma manager
├── pdf_processor.py   # PDF → LangChain Documents
└── requirements.txt

---